The Duality of the AI Build-out
There is a funny thing happening right now where every tech company is simultaneously building an AI agent into their product, and building their product into AI chat apps like ChatGPT and Claude.
Nobody knows which paradigm will win, so everyone is doing both. Many software products now have a specialized agent built in, and an MCP plugin for general purpose agents.
Should I use cal.com over MCP, or pay for their AI credits? Or both?
As a developer, both paths have clear advantages and disadvantages. A specialized agent is probably going to be more effective and token efficient at your product’s particular tasks than a general purpose chat app can be. Having it present in your product’s UI keeps users in your app, and you can build a more powerful integration when you’re not limited by the MCP plugin release process, iframes, and lack of system prompt control. You own the full stack, and every pixel of the UI. An integrated, specialized agent is probably a better experience for your users overall, at least on the dimensions of speed and execution quality.
On the other hand, you need to pass the cost of inference through to your users, at API prices. It also fragments the user experience: people end up juggling separate AI agents across different products, none of which share context.
Having an MCP plugin solves this, because you don’t have to pay for the inference. Your plugin is one of many that ChatGPT and Claude can connect to, so in some ways your product is probably going to be more useful as a component people can build workflows with. There is also the obvious fact that being available in ChatGPT and Claude exposes you to a massive existing installed base. These are the new app store. And people already spend a lot of time inside these chat apps, so the friction to using your product is probably lower.
But there is this lack of control, and “separateness” that you have to contend with. Either you embed a mini version of your product into the chat app as a widget, or you expect users to just not use your product’s UI or perhaps use the two side-by-side.
Having both of these architectures at once is also confusing to users because the context is fragmented - you can talk to your Notion agent one day, then ask ChatGPT about that thing you did in Notion later and it will have no knowledge of that conversation. It’s easy to get mixed up with agent fatigue.
OpenAI offers a hybrid model with “Sign in with ChatGPT”, which lets apps use subscription inference in their native AI features. This feels to me like the most user-friendly model if it becomes generally available. I suppose here an AI lab has to decide what they value more - API margins on tokens, or the ubiquity and lock-in that come with being an auth system like “Sign in with Google”.
Longer term there’s also the local-first architecture, where people have on-device inference good enough that software products’ native agents can just use that over something like the Prompt API. This could very well end up winning overall.
This is a huge architectural change to the internet which for now is completely unsettled, and I’m really curious where it all ends up.